Ask Claude where mobile customers get stuck

Liquid brings voice, vision, and decisions to small local models

Liquid AI released LFM3, an open-weight family for specialized on-device apps. You can customize the models on a single GPU rather than build around a hosted general-purpose model.

• The 1.7B text model supports a 256K context window for working with long inputs.
• The vision model accepts video input; the audio model supports tool calls, voice chat, and reasoning in 18 languages.
• Two decision models return typed outputs and calibrated probabilities in a single pass, with audio input available in the Omni version.

What it means Try it if you’re building a customized on-device voice or vision app. Start with the model that matches your input, rather than the whole multimodal suite.

Source: liquidai

Supabase previews agents that outlast an Edge Function

Supabase Compute puts agent sandboxes, background jobs, and always-on services inside your Supabase project. It gives work such as an hour-long agent run or a Python API a place to execute alongside your database.

• Each workload runs in its own isolated microVM and inherits your project’s authentication, roles, and permissions.
• You can restrict the external endpoints and ports a workload can reach; logs appear in your dashboard.
• Compute is in private alpha, with access through a waitlist.

What it means Join the waitlist if your Supabase app needs jobs that don’t fit in an Edge Function. This isn’t an option for an immediate production migration.

Source: supabase

Quick hits

• Google previews Genie 4 with real-time scene controls, native audio, and interactive dialogue; trusted testers get access first, with Ultra and developer access coming later.
• Anthropic disclosed its own drug-discovery programs and early results from six partnerships, expanding its scientific work into real drug programs that remain mostly preclinical.
• An Anthropic study of 50 analysts found their choices matched agents’ randomly assigned preferences 86% of the time across four analytical decisions, challenging assumptions about independent human review.
• Replit Cloud Agents turn plain-language requests into tested pull requests, with engineers approving the final code before it ships.
• Anthropic’s expanded GSA agreement offers eligible US governments 40% off Claude products and API access, with no minimum spend; cloud-provider discounts follow in coming months.

GitHub repo of the day

Keep parallel coding agents in separate workspaces

ThinkRail’s README describes an interface for running parallel coding-agent sessions in separate Git branches and working folders. You can review file changes, send comments to an agent, and keep project knowledge in SPEC.md files beside the code. It offers desktop and browser interfaces and requires Git plus a connected model provider.

AI workflow of the day

How to find friction in your mobile booking flow

Shared by Noah Kagan (@noahkagan) · View original post ↗

If customers must scroll or tap through too much before booking, this Claude prompt gives you a focused fix list. Noah Kagan reports that TidyCal tested its changes on 63K visitors, with bookings rising from 12.2% to 13.0%.